1

29. Frame Semantics 2

Jean Mark Gawron 3
San Diego State University 4
5
6
Department of Linguistics
5500 Campanile Drive
San Diego State University
San Diego, CA 92182-7727
(619)-594-0252
7
Keywords: semantics, frames, lexicon 8
19. Frame semantics 9
1. Introduction 10
2. Fillmorean frames 11
3. Related conceptions 12
4. Events, proling, and perspectivalization 13
5. Lexicography 14
6. Discourse understanding 15
7. Conclusion 16
8. References 17
Abstract 18
Frames are conceptual structures that provide context for elements of in- 19
terpretation; their primary role in an account of text understanding is to 20
explain how our text interpretations can leap far beyond what the text lit- 21
erally says. The present article explores the role of frames in providing a 22
principled account of the openness and richness of word-meanings, distin- 23
guishing a frame-based account from classical approaches, such as accounts 24
based on conceptual primitives, lexical elds, and connotation, and showing 25
how they can play a role in the account of how word meaning interacts with 26
syntactic valence. 27
1
28
For there exists a great chasm between those, on the one side, who
relate everything to a single central vision, one system more or less
coherent or articulate, in terms of which they understand, think and
feel a single, universal, organizing principle in terms of which alone
all that they are and say has signicance and, on the other side,
those who pursue many ends, often unrelated and even contradictory,
connected, if at all, only in some de facto way, for some psychological
or physiological cause, related by no moral or aesthetic principle.
29
Berlin (1957: 1), cited by Minsky (1975) 30
1. Introduction 31
Two properties of word meanings contribute mightily to the diculty of 32
providing a systematic account. 33
One is the openness of word meanings. The variety of word meanings 34
is the variety of human experience. Consider dening words such as Tues- 35
day, barber, alimony, seminal, amputate, and brittle. One needs to make 36
reference to diverse practices, processes, and objects in the social and phys- 37
ical world: repeatable calendar events, grooming and hair, marriage and 38
divorce, discourse about concepts and theories, and events of breaking. Be- 39
2
fore this seemingly endless diversity, semanticists have in the past stopped 40
short, excluding it from the semantic enterprise, and attempting to draw a 41
line between a small linguistically signicant set of primitive concepts and 42
the openness of the lexicon. 43
The other problem is the closely related problem of the richness of word 44
meanings. Words are hard to dene, not so much because they invoke 45
ne content specic distinctions, but because they invoke vast amounts of 46
background information. The concept of buying presupposes the complex 47
social fact of a commercial transaction. The concept of alimony presupposes 48
the complex social fact of divorce, which in turn presupposes the complex 49
social fact of marriage. Richness, too, has inspired semanticists simply to 50
stop, to draw a line, saying exact denitions of concepts do not matter for 51
theoretical purposes. 52
This boundary-drawing strategy, providing a response if not an answer to 53
the problems of richness and openness, deserves some comment. As linguistic 54
semanticists, the story goes, our job is to account for systematic, structurally 55
signicant properties of meaning. This includes: 56
(1) a. the kinds of syntactic constructions lexical meanings are compat- 57
ible with. 58
i. the kinds of participants that become subjects and objects 59
3
ii. regular semantic patterns of oblique markings and valence al- 60
ternations 61
b. Regular patterns of inference licensed by category, syntactic con- 62
struction or closed class lexical item. 63
The idea is to carve o that part of semantics necessary for knowing and 64
using the syntactic patterns of the language. To do this sort of work, we do 65
not need to pay attention to every conceptually possible distinction. Instead 66
we need a small set of semantic primitives that make the distinctions that 67
linguistically matter; what is left over can be dealt with using some open 68
class of predicates or features whose internal details are not of concern. 69
Jackendo (1990) is a good example of this kind of approach. The genera- 70
tive semantics program, especially as outlined in Lako (1972), is another. 71
Dowty (1979) has many of the same features, but in places expresses doubts 72
that the program can be completely carried out. The kind of analysis I 73
have in mind can be exemplied through Dowtys generative-semantics-like 74
analysis of causatives like break.tr (transitive break): 75
(2) a. John broke the glass. 76
b. do(John, cause(become(broken(glass)))) 77
Here the predicates in capitals (do, cause, become) are from the inventory 78
of linguistically signicant primitives, and the lower case predicates (broken, 79
4
glass) are from the open class predicates whose internal structure does not 80
matter. At most we need to know that one expresses a state (broken) and the 81
other a kind (glass). The details beyond that are linguistically insignicant. 82
Of course there are dierences in truth-conditions between states like broken 83
and dead, but these have only minor selectional eects on the causative 84
inchoatives created from them (break.tr = do ... cause become broken 85
and kill = do ... cause become dead). I will refer to this view of lexical 86
semantics as the classical view. 87
In this paper I wish to consider a view of semantics in general and lexi- 88
cal semantics in particular that is quite at odds with this classical picture: 89
frame semantics (Fillmore 1975, 1977b, 1978, 1982, 1985). Someone wish- 90
ing to contest the classical picture has two options: rst, contend that the 91
wrong kinds of questions are being asked; second, argue that the program 92
as outlined is not very well-suited to attaining its goals. As we shall see, 93
both kinds of objection motivate frame semantics. 94
5
2. Fillmorean frames 95
2.1. Motivations 96
The version of frame semantics I will present here is largely the brainchild of 97
Charles J. Fillmore. Although frame semantics has sprouted o in a number 98
of directions and been applied to a number of problems, I will limit the 99
present discussion in two ways: First I will conne myself largely to eshing 100
out the Fillmorean picture; second, I will conne myself mostly to questions 101
of the lexicon, lexicography, and the lexicon-syntax interface, leaving for 102
other work questions of discourse and text understanding to which frames 103
are also relevant. I will briey consider the dierent roles frames play in the 104
account of sign meaning and discourse interpretation. 105
Although Fillmore has had many interesting things to say about the 106
kinds of problems listed in (1) in early and late works on Case Grammar, the 107
primary motivations given in Fillmore (1982, 1985) focus on frame semantics 108
as a contribution to a theory of text understanding. Consider for example, 109
the very dierent scenes evoked by the following pair of sentences, discussed 110
in Fillmore (1985): 111
(3) a. I cant wait to be on the ground again. 112
b. I cant wait to be on land again. 113
6
Sentence (3a) evokes a speaker who is in the air (on a plane), sentence (3b) 114
a speaker who is at sea (on a ship). This contrast is tied to some dierence 115
between the words land and ground, yet, on the face of it, land and ground 116
denote very similar things. Fillmore would say land is understood within 117
a conceptual frame of sea travel, and within that frame it is opposed to 118
sea, while ground is understood within a conceptual frame of air travel, and 119
within that frame, it is opposed to air. Thus we can explain something that 120
is very dicult to explain in terms what the words in the sentence denote 121
by investigating the conceptual background against which the relevant word 122
senses are dened. That conceptual background is what Fillmore calls a 123
frame. 124
Frames are conceptual structures that provide context for elements of 125
interpretation; their primary role in an account of text understanding is to 126
explain how our text interpretations can (validly) leap far beyond what the 127
text literally says. Frames can be introduced into interpretation in a variety 128
of ways. They may be directly tied to word senses as in the example of land 129
and ground or they may be introduced by patterns among the facts the text 130
establishes. To use another example of Fillmores (1985: 232): 131
(4) We never open our presents until morning. 132
This sentence evokes the Christmas frame by describing a situation that 133
7
matches salient facts of Christmas practice, even though no word in it is 134
specic to Christmas. If in fact the Christmas frame is the right one, that 135
evocation makes a signicant contribution to the understanding of the sur- 136
rounding text. 137
Frames are motivated not just by words, then, but by stereotypes about 138
customs, practices, institutions, and games. Moreover, the kinds of cog- 139
nitive structures Fillmore has in mind have been proposed by a variety of 140
researchers for a variety of purposes. Fillmore has adopted the terminol- 141
ogy of AI researcher Minsky (1975) in calling them frames, but schemata in 142
psychology (Bartlett 1932, Rumelhart 1980) are getting at something very 143
similar, as are scripts (Schank & Abelson 1977), cognitive models (Lako 144
1983), experiential gestalts (Lako & Johnson 1980), the base (as opposed to 145
the prole) (Langacker 1984), and Fillmores own notion of scene (Fillmore 146
1976, 1977a). More recently, in articulating a simulation view of conceptual 147
processing, Barsalou (1992, 1999) has proposed that object conceptualiza- 148
tion is processed through simulators of objects linked to components of a 149
variety of situation memories; one consequence is that objects may acti- 150
vate components from dierent situations in dierent perceptual contexts. 151
In this theory, too, then, conceptualization is framed against a background 152
with components that help provide an interpretation for scenes or objects. 153
8
For more discussion, see article 108 (Kelter & Kaup) Conceptual knowledge, 154
categorization, and meaning. 155
As an approach to word meanings specically, the starting point for 156
frame semantics is that the lexical semantics problems of openness and 157
richness are connected. Openness depends on richness. Openness does not 158
mean lack of structure. In fact, it presupposes structure. Most concepts are 159
interpretable or understandable or denable only against the background 160
of other concepts. Many backgrounds are rich enough to dene a cluster 161
of concepts, in particular, a cluster of words. These backgrounds are the 162
frames. Thus because words are networked together through their shared 163
backgrounds, frames can provide an organizing principle for the openness of 164
the lexicon. 165
Consider one of the examples already discussed, discussed in Fillmore 166
(1982). The concept of alimony depends on the concept of divorce. The 167
concept of divorce in turn depends on the concept of marriage. The depen- 168
dency is denitional. Unless you dene what a marriage is, you cant dene 169
what a divorce is. Unless you dene what a divorce is, you cant dene 170
what alimony is. Thus there is a very real sense in which the dependencies 171
we are describing move us toward simpler concepts. Notice, however, that 172
the dependency is leading in a dierent direction than an analysis that de- 173
9
composes meanings into a small set of primitives like cause and become. 174
Instead of leading to concepts of increasing generality and abstractness, we 175
are being led to dene the situations or circumstances which provide the 176
necessary background for the concepts we are describing. The concepts of 177
marriage and divorce are equally specic, but the institution of marriage 178
provides the necessary background for the institution of divorce. 179
Or consider the complex subject of Tuesdays (Fillmore 1985). We live in 180
a world of cyclic events. Seasons come and go and then return. This leads 181
to a cyclic calendar which divides time up into repeating intervals, which are 182
divided up further. Years are divided into months, which are divided into 183
weeks, which are divided into days, which have cyclic names. Each week 184
has a Sunday, a Monday, a Tuesday, and so on. Dening Tuesday entails 185
dening the notion of a cyclic calendar. Knowing the word Tuesday may not 186
entail knowing the word Sunday, but it does entail understanding at least 187
the concept of a week and a day and their relation, and that each week has 188
exactly one Tuesday. 189
We thus have words and background concepts. We will call the back- 190
ground concept the frame. Now the idea of a frame begins to have some 191
lexical semantic bite with the observation that a single concept may provide 192
the background for a set of words. Thus the concept of marriage provides 193
10
the background for words/suxes/phrases such as bride, groom, marriage, 194
wedding, divorce, -in-law, elope, ancee, best man, maid-of-honor, honey- 195
moon, husband, and wife, as well as a variety of basic kinship terms omitted 196
here for reasons of space. The concept of calendar cycle provides the 197
frame for lexical items such as week, month, year, season, Sunday, ..., Sat- 198
urday, January, ..., December, day, night, morning, and afternoon. Notice 199
that a concept once dened may provide the background frame for further 200
concepts. Thus, divorce itself provides the background frame for lexical 201
items such as alimony, divorce, divorce court, divorce attorney, ex-husband, 202
and ex-wife. 203
In sum, a frame may organize a vocabulary domain: 204
Borrowing from the language of gestalt psychology we could say 205
that the assumed background of knowledge and practices the 206
complex frame behind this vocabulary domain stands as a 207
common ground to the gure representable by any of the indi- 208
vidual words. 209
[Words belonging to a frame] are lexical representatives of some 210
single coherent schematization of experience or knowledge. 211
Fillmore (1985: 223) 212
11
Now a premise of frame semantics is that the relation between lexical 213
items and frames is open ended. Thus one way in which the openness of 214
the lexicon manifests itself is in building concepts in unpredictable ways 215
against the backdrop of other concepts. The concept of marriage seems to 216
be universal or near-universal in human culture. The concept of alimony 217
is not. No doubt concepts sometimes pop into the lexicon along with their 218
dening frames (perhaps satellite is an example), but the usual case is to 219
try to build them up out of some existing frame (Thus horseless carriage 220
leading to car is the more usual model). 221
Summing up: openness does not mean structurelessness. Concepts and 222
their related words have certain unidirectional backgrounding relations that 223
frames capture. 224
12
(5) Words Frames
bride, groom, marriage, wedding, divorce, -in-
law, elope, ancee, best man, maid-of-honor,
honeymoon, husband, wife
marriage
alimony, divorce court, divorce attorney, ex-
husband, and ex-wife
divorce
week, month, year, Sunday, ..., Saturday, Jan-
uary, ..., December, morning, afternoon
calendar cycle
freezing, cold, cool, tepid, lukewarm, warm,
hot, temperature, thermometer
temperature
225
All of this obviously points in exactly the opposite direction from the 226
classical view, a few salient primitives, a hard distinction between linguistic 227
and encyclopedic, and a large uninvestigated class of open class predicates. 228
But from the other direction, support for the classical view has been 229
eroding even among those whose concerns have primarily departed from 230
the problems in (1) such as Levin (1993) or from classic lexical semantic 231
problems like polysemy (Pustejovsky 1995). 232
Consider the kind of problem Levin (1993) discusses in her seminal study 233
of of English verb classes. A theory that does not posit a systematic dier- 234
ence between the broken state of the verb break in (2) and the dead state in 235
13
the decomposition of kill cannot account for the following contrast: 236
(6) a. John broke the glass against the wall. 237
b. # John killed the cockroach against the wall. 238
Nor can it account for the fact that verbs in some sense close in meaning to 239
break (shatter, smash, crack, atten) will follow pattern (a), while verbs in 240
some sense close to kill will follow pattern (b) (strangle, murder, smother, 241
and drown). The generalization at issue is (roughly) that state change or 242
directed action verbs whose eect is commonly achieved by moving one 243
object against another will allow pattern (a) when the object whose state 244
is changed or potentially changed is direct object. Other examples are hit, 245
knock, rap, bang, and slam. None of the kill-type verbs t the bill. 246
Thus if valence patterns are part of what is to be explained, then a lan- 247
guage like English, with its rich inventory of prepositions and situationally 248
specic constructions (see for example the pattern lists in Levin 1993), will 249
require reference to a large inventory of concepts. It is dicult to see how 250
a principled line between open class and closed class concepts can be drawn 251
in carrying out this program. It is clear for example, that Levins verbs of 252
contact, which include the verbs like hit and slap discussed above, overlap 253
signicantly with the verbs list for the impact frame in FrameNet, a large 254
computational instantiation of the ideas of frame semantics (Fillmore & 255
14
Atkins 1994, Baker, Fillmore & Lowe 1998, Fillmore & Atkins 1998, Baker 256
& Fillmore 2001, Boas 2001, 2005, Chang, Narayanan & Petruck 2002a, 257
2002b). At last count the NSF Framenet project (Fillmore & Baker 2000) 258
which is building a frame lexicon for English had over 800 frames for about 259
4500 words. Thus the problems of openness and richness arise whether one 260
starts from text understanding or from syntax/semantics interface. 261
2.2. Basic Tools 262
We have thus far focused on the role of frames in a theory of word meanings. 263
Note that nothing in particular hangs on the notion word. Frames may also 264
have a conventional connection to a simple syntactic constructions or idiom; 265
give someone the slip probably belongs to the same frame as elude. Or 266
they may be tied to more complex constructions such as the Comparative 267
Correlative (cf. article 86 (Kay & Michaelis) Constructional meaning). 268
(7) The more I drink the better you look. 269
This construction has two slots requiring properties of quantity or de- 270
gree. The same issues of background situation and proled participants 271
arise whether the linguistic exponent is a word or construction. The term 272
sign, used in exactly the same sense as it is used by construction grammar- 273
ians, will serve here as well. 274
15
As a theory of the conventional association of schematized situations and 275
linguistic exponents, then, frame semantics makes the assumption that there 276
is always some background knowledge relative to which linguistic elements 277
do some proling, and relative to which they are dened. Two ideas are 278
central: 279
1. a background concept 280
2. a set of signs including all the words and constructions that utilize this 281
conceptual background. 282
Two other important frame theoretic concepts are frame elements and pro- 283
ling. 284
Thus far in introducing frames I have emphasized what might be called 285
the modularity of knowledge. Our knowledge of the the world can usefully 286
be divided up into concrete chunks. Equally important to the Fillmorian 287
conception of frames is the integrating function of frames. That is, frames 288
provide us with the means to integrate with other frames in context to 289
produce coherent wholes. For this function, the crucial concept is the notion 290
of a frame element (Fillmore & Baker 2000). A frame element is simply a 291
regular participant, feature, or attribute of the kind of situation described 292
by a frame. Thus, frame elements of the wedding frame will include the 293
16
husband, wife, wedding ceremony, wedding date, best man and maid of 294
honor, for example. Frame elements need not be obligatory; one may have a 295
wedding without a best man; but they need to be regular recurring features. 296
Thus, frames have slots, replaceable elements. This means that frames 297
can be linked to to other frames by sharing participants or even by being 298
participants in other frames. They can be components of an interpretation. 299
In frame semantics, all word meanings are relativized to frames. But 300
dierent words select dierent aspects of the background to prole (we use 301
the terminology in Langacker 1984). Sometimes aspects proled by dierent 302
words are mutually exclusive parts of the circumstances, such as the husband 303
and wife in the marriage frame, but sometimes word meanings dier not in 304
what they prole, but in how they prole it. In such cases, I will say words 305
dier in perspective (Fillmore 1977a). I will use Fillmores much-discussed 306
commercial event example (Fillmore 1976) to illustrate: 307
(8) a. John sold the book to Mary for $100. 308
b. Mary bought the book from John for $100. 309
c. Mary paid John $100 for the book. 310
Verbs like buy, sell, pay, have as background the concept of a commercial 311
transaction, an event in which a buyer, gives money to a seller in exchange for 312
some goods. Now because the transaction is an exchange it can be thought of 313
17
as containing what Fillmore calls two subscenes: a goods transfer, in which 314
the goods is transferred from the seller to the buyer, and a money transfer, 315
in which the money is transferred from the buyer to the seller. Here it is 316
natural to say that English has as a valence realization option for transfers 317
of possession one in which the object being transferred from one possessor to 318
another is realized as direct object. Thus verbs proling the money transfer 319
will make the money the direct object (pay and collect) and verbs proling 320
the goods transfer will make the goods the direct object (buy and sell). Then 321
the dierence between these verb pairs can be chalked up to what is proled. 322
But what about the dierence between buy and sell? By hypothesis, 323
both verbs prole a goods transfer, but in one case the buyer is subject and 324
in another the seller is. Perhaps this is just an arbitrary choice. This is in 325
some sense what the thematic role theory of Dowty (1991) says: Since (8a) 326
and (8b) are mutually entailing, there can be no semantic account of the 327
choice of subject. 328
In frame semantics, however, we may attempt to describe the facts as 329
follows: in the case of buy the buyer is viewed as (perspectivalized as) agent, 330
in the case of sell, the seller is. There are two advantages to this descrip- 331
tion. First, it allows us to preserve a principle assumed by a number of 332
linguists, that cross-linguistically agents must be subjects. Second, it allows 333
18
us to interpret certain adjuncts that enter into special relations with agents: 334
instrumentals, benefactives, and purpose clauses. 335
(9) a. John bought the book from Mary with/for his last pay check. 336
[both with and for allow the reading on which the pay check pro- 337
vides the funds for the purchase.] 338
b. Mary sold the book to John ?with/for his last paycheck. [Only for 339
allows the reading on which the pay check provides the funds.] 340
c. John bought the house from Sue for Mary. [allows reading on 341
which Mary is ultimate owner, disallows the reading on which 342
Mary is seller and Sue is sellers agent] 343
d. Sue sold the house to John for Mary. [allows reading on which 344
Mary is seller and Sue is sellers agent; disallows reading on which 345
Mary is ultimate owner.] 346
e. John bought the house from Sue to evade taxes/as a tax dodge. 347
[tax benet is Johns] 348
f. Sue sold the house to John to evade taxes/as a tax dodge. [tax 349
benet is Sues] 350
But what does it mean to say that a verb takes a perspective which 351
views a particular participant as an agent? The facts are, after all, that 352
both the buyer and the seller are agents; they have all the entailment prop- 353
19
erties that characterize what we typically call agents; and this, Dowtys 354
theory of thematic roles tells us, is why verbs like buy and sell can co-exist. 355
I will have more to say on this point in section 4; for the moment I will 356
conne myself to the following general observation on what frame semantics 357
allows: What is proled and what is left out is not determined by the en- 358
tailment facts of its frame. Complex variations are possible. For example, 359
as Fillmore observes, the commercial transaction frame is associated 360
with verbs that have no natural way of realizing the seller: 361
(10) John spent $100 on that book. 362
Nothing in the valence marking of the verb spend suggests that what is being 363
proled here is a possession transfer; neither the double object construction, 364
nor from nor to is possible for marking a core commercial transaction 365
participant. Rather the pattern seems to be the one available for what one 366
might call resource consumption verbs like waste, lose, use (up), and blow. 367
In this proling, there is no room for a seller. Given that such variation 368
in what is proled is allowed, the idea that the agenthood of a participant 369
might be part of whats included or left out does not seem so far-fetched. 370
As I will argue in section 4, the inclusion of events into the semantics can 371
help us make semantic sense of what abstractions like this might mean. 372
20
These considerations argue that there can be more than one frame back- 373
grounding a single word meaning; for example, concepts of commercial event, 374
possession transfer, and agentivity simultaneously dene buy. A somewhat 375
dierent but related issue is the issue of event structure. There is strong ev- 376
idence cross-linguistically at least in the form of productive word-formation 377
processes that some verbs for example, causatives represent complex 378
events that can only be expressed through a combination of two frames 379
with a very specic semantics. Article 107 (Libben & de Almeida) Mental 380
lexicon examines additional psycholinguistic evidence for complex events. 381
So it appears that a word meaning can simultaneously invoke a congura- 382
tion of frames, with particulars of the conguration sometimes spelled out 383
morphologically. 384
The idea that any word meaning exploits a background is of use in the 385
account of polysemy. Dierent senses will in general involve relativization 386
to dierent frames. As a very simple example, consider the use of spend in 387
the following sentence: 388
(11) John spent 10 minutes xing his watch. 389
How are we to describe the relationship of the use of spend in this example, 390
which basically describes a watch xing event, with that in (10), which 391
describes a commercial transaction? One way is to say that one sense 392
21
involves the commercial transaction, and another involves a frame we 393
might call action duration which relates actions to their duration, a frame 394
that would also be invoked by durative uses of for. A counter-proposal is 395
that there is one sense here, which involves an actor using up a resource. 396
But such a proposal runs up against the problem that spend really has rather 397
odd disjunctive selection restrictions: 398
(12) John spent 30 packs of cigarettes that afternoon. 399
Sentence (12) is odd except perhaps in a context (such as a prison or board- 400
ing school) where cigarette packs have become a fungible medium of ex- 401
change; what it cannot mean is that John simply used up the cigarettes (by 402
smoking them, for example). The point is that a single general resource 403
consumption meaning ought to freely allow resources other than time and 404
money, so a single resource consumption sense does not correctly describe 405
the readings available for (12); however, a sense invoking a commercial 406
transaction frame constrained to very specic circumstances does. Note 407
also, that the fact that 30 packs of cigarettes can be the money participant 408
in the right context is naturally accommodated. The right constraint on the 409
money participant is not that it be cash (for which Visa and Mastercard can 410
be thankful), but that it be a fungible medium of exchange. 411
Summarizing: 412
22
1. Frames are motivated primarily by issues of understanding and con- 413
verge with various schema-like conceptions advanced by cognitive psy- 414
chologists, AI researchers, and cognitive linguists. They are experien- 415
tially coherent backgrounds with variable components that allow us to 416
organize families of concepts. 417
2. The concept of frames has far reaching consequences when applied to 418
lexical semantics, because a single frame can provide the organizing 419
background for a set of words. Thus frames can provide an organizing 420
principle for a rich open lexicon. FrameNet is an embodiment of these 421
ideas. 422
3. In proposing an account of lexical semantics rich enough for a theory of 423
understanding, frame semantics converges with other lexical semantic 424
research which has been bringing to bear a richer set of concepts on 425
problems of the syntax semantics interface. 426
Having sketched the basic idea, I want in the next two sections to briey 427
contrast the notion frame with two other ideas that have played a major 428
role in semantics, the idea of a relation, as incorporated via set theory and 429
predicate logic into semantics, and the idea of a lexical eld. 430
23
3. Related conceptions 431
In this section I compare the idea of frames with two other concepts of major 432
importance in theories of lexical semantics, relations and lexical elds. The 433
comparison oers the opportunity to develop some other key ideas of frame 434
semantics, including proling and saliency. 435
3.1. Frames versus relations: proling and saliency 436
Words (most verbs, some nouns, arguably all degreeable adjectives) describe 437
relations in the world. Love and hate are relations between animate experi- 438
encers and objects. The verb believe describes a relation between an animate 439
experiencer and a proposition. These are commonplace views among philso- 440
phers of language, semanticists, and syntacticians, and they have provided 441
the basis for much fruitful work. Where do frames t in? 442
For Fillmore, frames describe the factual basis for relations. In this 443
sense they are pre-relational. To illustrate, Fillmore (1985) cites Mills 444
(1847) discussion of the words father and son. Although there is a single 445
history of events which establishes both the father- and the son- relation, 446
the words father and son pick out dierent entities in the world. In Mills 447
terminology, the words denote dierent things, but connote a single thing, 448
the shared history. This history, which Mill calls the fundamentum relatio- 449
24
nis (the foundation of the relation), determines that the two relations bear 450
a xed structural relation to each other. It is the idea of a determinate 451
structure for a set relations that Fillmore likens to the idea of a frame. 452
Thus, a frame denes not a single relation but, minimally, a structured 453
set of relations. 454
This conception allows for a natural description not just of pairs of words 455
like father and son, but also of single words which do not in fact settle 456
on a particular relation. Consider the verb risk, discussed in Fillmore & 457
Atkins (1998), which seems to allow a range of participants into a single 458
grammatical slot. For example, 459
(13) Joan risked
_

_
a. censure.
b. her car.
c. a trip down the advanced ski slope.
460
The risk frame has at least 3 distinct participants, (a) the bad thing that 461
may happen, (b) the valued thing that may be lost, and (c) the activity 462
that may cause the bad thing to happen. All can be realized in the direct 463
object position, as (13) shows. Since there are three distinct relations here, a 464
theory that identies lexical meanings with relations needs to say there are 3 465
meanings as well. Frame semantics would describe this as one frame allowing 466
3 distinct prolings. It is the structure of the frame together with the 467
25
proling options the language makes available which makes the 3 alternatives 468
possible. 469
Other verbs with a similar indeterminacy of participant are copy, collide, 470
and mix: 471
(14) a. Sue copied her costume (from a lm poster). 472
b. Sue copied the lm poster. 473
c. The truck and the car collided. 474
d. The truck collided with the car. 475
e. John mixed the soup. 476
f. John mixed the paste into the soup. 477
g. John mixed the paste and the our. 478
In each of these cases the natural frame semantics account would be to 479
say the frame remains constant while the prolings or perspective changes. 480
Thus, under a frame semantics approach, verbal valence alternations are to 481
be expected, and the possibility of such alternations provides motivation for 482
the idea of a background frame with a range of participants and a range of 483
proling options. 484
Now on a theory in which senses are relations, all the verbs in (14) must 485
have dierent senses. This is, for example, because the arguments in (14a) 486
and (14b) ll dierent roles. Frame semantics allows another option. We 487
26
can say the same verb sense is used in both cases. The dierences in inter- 488
pretation arise because of dierences in proling and perspectivalization. 489
3.2. Frames versus lexical elds 490
Because frames dene lexical sets, it is useful to contrast the concept of 491
frames with an earlier body of lexical semantic work which takes as central 492
the identication of lexical sets. This work develops the idea of lexical elds 493
(Weisgerber 1962, Coseriu 1967, Trier 1971, Geckeler 1971, Lehrer & Kittay 494
1992). Lexical elds dene sets of lexical items in mutually dening relations, 495
in other words, lexical semantic paradigms. The classic example of a lexical 496
eld is the set of German labels used for evaluating student performance 497
(Weisgerber 1962: 99): 498
(15) sehr gut, gut, gen ugend and mangelhaft 499
The terms are mutually dening because the signicance of a single evalu- 500
ation obviously depends on knowing the entire set and the relations of the 501
terms in the set. Thus gut means one thing in a school system with the 4 502
possibilities in (15) and quite another if the possibilities are: 503
(16) sehr gut, gut, befriedigend ausreichend, mangelhaft and ungen ugend 504
Fillmore also cites the example of the tourist industry use of the term rst 505
class in their categorization of hotels; to many travelers, rst class sounds 506
27
pretty good; in fact, the top ranked class of hotels is luxury and rst class 507
is fourth from the top. The misunderstanding here seems exactly like a case 508
of applying the wrong frame in the process of understanding. 509
Domains in which lexical elds have provided fruitful analyses include 510
color, temperature, furniture and artifacts, kinship relations, intelligence, 511
livestock, and terrain features (Fillmore 1985: 227). 512
The general hypothesis of lexical eld theory is that the lexicon can 513
be carved up into a number of (sometimes overlapping) lexical sets, each 514
of which functions as a closed system. To this extent, there is agreement 515
with the conception of frames, and in fact, the lexical sets associated with 516
frames can include lexemes in paradigmatic, mutually dening relations. 517
For example, we identied the temperature frame in section 2, and this 518
includes the lexical eld of temperature words like cold, cool, lukewarm, 519
warm, and hot. 520
However, the idea of a frame is distinct from the idea of a lexical eld. 521
To start with, the idea of a one-word lexical eld is incoherent: How can 522
a word have a function in a eld in which there is nothing for it to be 523
opposed to? However, there is no inherent diculty with the idea of a one- 524
word frame. Fillmore (1985) cites the example of hypotenuse, which requires 525
for its background the concept of a right triangle. There appear to be no 526
28
other English lexical items specic to right triangles (the term leg in the 527
relevant sense seems to apply to triangle sides in general); and that is neither 528
surprising nor problematic. The notion mutually dening is not necessary 529
for lexical frame sets because words in frames are dened in contrast to or in 530
terms of the frame alone. The frame, not its lexical instantiations, provides 531
the background necessary to identify a semantic function. The primitive 532
notion is not dened in opposition to but proled from the background of. 533
A second way in which frames dier from lexical elds is that, even when 534
there is more than one word, there is no requirement that words in the set 535
function in paradigmatic opposition to one another. Thus the temper- 536
ature frame cited above also contains the noun temperature, just as the 537
height frame containing polar adjectives like tall and short will contain the 538
noun height. 539
Thirdly, because of the notion of mutual denition, lexical elds come 540
with strict criteria of individuation. In contrast, as we saw in section 2, 541
frames of arbitrary specicity make sense. Thus, we have very general 542
frames of temperature and height. But we also have a set of specic 543
frames that recover the traditional mutually dening sets that preoccupied 544
lexical eld theorists, a specialization of height that includes just the polar 545
adjectives, a specialization of temperature that includes just the set cold, 546
29
cool, warm, hot, and so on. This level of specicity in fact roughly describes 547
the granularity of FrameNet. 548
3.3. Minskian frames 549
As described in Fillmore (1982), the term frame is borrowed from Marvin 550
Minsky. It will be useful before tackling the question of how proling and 551
perspectivalization work to take a closer look at this precursor. 552
In Minskys original frames paper (Minsky 1975), frames were put forth 553
as a solution to the problem of scene interpretation in vision. Minskys pro- 554
posal was in reaction to those who, like the Gestalt theorists (Koka 1963), 555
viewed scene perception as a single holistic process governed by principles 556
similar to those at work in electric elds. Minsky thought scenes were as- 557
sembled in independent chunks, constituent by constituent, in a series of 558
steps involving interpretation and integration. To describe this process, a 559
model factoring the visual eld into a number of discrete chunks, each with 560
its own model of change with its own discrete phases, was needed. 561
A frame was thus a dynamic model of some specic kind object with 562
specic participants and parameters. The model had built-in expectations 563
about ways in which the object could change, either in time or as a viewers 564
perspective on it changed, formalized as operations mapping old frame states 565
30
C
A
B
E
cube
invis left
top right invis
D A E B C
fg
fg
fg
Fig. 29.1: View of cube together with simplied cube frame representing that
view. Links marked fg lead to foregrounded slots; slots marked invis
are backgrounded. Faces D and C are out of view.
to new frame states. A frame also included a set of participants whose sta- 566
tus changed under these operations; those moving into certain distinguished 567
slots are foregrounded. Thus, for example, in the simplied version of Min- 568
skys cube frame, shown before and after a rotation in gures 29.1 and 29.2, 569
a frame state encodes a particular view of a cube and the participants are 570
cube faces. One possible operation is a rotation of the cube, dened to place 571
new faces in certain view-slots, and move old faces out and possibly out of 572
view. The faces that end up in view are the foregrounded participants of the 573
resulting frame state. Thus the cube frame oers the tools for representing 574
particular views or perspectives on a cube, together with the operations that 575
31
C
A
D
E
cube
invis left
top right invis
D A E B C
fg
fg
fg
Fig. 29.2: Cube frame after counterclockwise rotation. Faces D and A are
now foregrounded, B has moved out of view.
may connect them in time. 576
Fillmores innovation, then, was to apply this Minskian idea in the do- 577
main of word meaning, importing not only the idea of chunked modular 578
knowledge units, but also the idea of operations that take perspectives on 579
such chunks. I used the terms proling and perspectivalization to describe 580
such operations in section 2. Although Fillmore himself does not attempt a 581
formalization of these operations, I believe it is possible to clearly describe 582
what is at issue using some ideas from event semantics (Davidson 1967, 1980, 583
Parsons 1990), building on the event-based approach to frames in Gawron 584
(1983). 585
32
4. Events, proling, and perspectivalization 586
To spell out a bit better how word senses might invoke multiple frames, let 587
us return to the case of the commercial transaction frame discussed 588
in section 2. The following development takes up and extends the ideas of 589
Gawron (1983). 590
A rather natural account of the interface between frames and composi- 591
tional semantics becomes available if we make use of neo-Davidsonian event- 592
semantics (Davidson 1967, 1980, Parsons 1990). On a neo-Davidsonian ac- 593
count, we have, as the schematic semantics for John bought the book on 594
sale: 595
e[buy(e)agent(e)=j patient(e)=b on-sale(e, b)] 596
We call e in the above representation the lexical event. 597
I assume that Fillmorean frames classify events. That is, there is such a 598
thing as a commercial transaction event. Further, I assume that lexi- 599
cal predicates like give and buy are predicates true of events. These lexical 600
events cannot be directly identied with Fillmorean frame events. Rather 601
the lexical events are perspectivalizations of Fillmorean frame events. Thus, 602
for example, buying will be associated with three events, one perspectivaliz- 603
33
ing event that is directly related to syntactic realization, a second proling 604
event that is a proling of a third commercial transaction (or Fill- 605
morean frame event). I will call this latter the circumstance event. Perspec- 606
tivalizing, proling, and circumstance events will be related by functions. 607
Borrowing the machinery of sorted logic (Carpenter 1992, Smolka 1992, 608
Rounds 1997), I will assume that all predicates are sorted; that is, it is 609
a property of predicates and relations that in all models, for any given 610
argument position, there is a sort of individuals for which that argument 611
position is dened. I will write sorts in boldface and predicates in roman. 612
(17) agent patient : agent patient truth-values
agent : agent patient animate
patient : agent patient entity
source : agent patient (entity)
goal : agent patient (entity)
613
These declarations just say, in roughly standard mathematical notation that 614
agent and patient are functions from one set to another. For example, the 615
rst declaration says that agent patient is a function from the set (sort) 616
to truth-values; the second says agent is a function from the set (sort) of 617
agent patient events to animates; patient from the set of agent patient 618
events to the set of things (the domain of entities). The parentheses in 619
34
the source and goal role denitions may be taken to mean that the role is 620
optional (or the function is partial). Not every agent patient event has a 621
source or a goal, but some do. 622
I assume the declarations (or axioms) in (17) are sucient to dene a very 623
simple kind of frame. The rst axiom denes a predicate agent patient 624
that is true of events of that sort; the rest dene a set of roles for that sort of 625
event. Thus a minimal frame is just an event sort dened for a set of roles. 626
I will call agent patient an argument frame because syntactic arguments 627
of a verb will need to directly link to the roles of argument frames (such as 628
agent and patient). We can represent this set of axioms as an attribute-value 629
matrix (AVM): 630
(18)
_

_
agent patient
agent animate
source entity
goal entity
patient entity
_

_
631
Henceforth I use AVM notation for readability, but the reader should bear 632
in mind that it is merely a shorthand for a set of axioms like those in (17), 633
constraining partial functions and relations on sorts. 634
I will call agent patient an argument frame because syntactic argu- 635
35
ments of a verb will need to directly link to the roles of argument frames 636
(such as agent and patient). The agent patient frame is very general, 637
too general to be of much semantic use. In order to use it a lexical item 638
must specify some circumstance frame in which participant roles are further 639
specied with further constraints. 640
The connection between an argument frame like agent patient and 641
simple circumstance frames can be illustrated through the example of the 642
possession transfer frame (related to verbs like give, get, take, receive, 643
acquire, bequeath, loan, and so on). Represented as an AVM, this is: 644
(19)
_

_
possession transfer
donor animate
possession entity
recipient animate
_

_
645
Now both give and acquire will be dened in terms of the possession trans- 646
fer frame, but give and acquire dier in that with give the donor becomes 647
subject and with acquire the recipient does. (Compare the dierence be- 648
tween buy and sell discussed in section 2.2.) 649
We will account for this dierence by saying that give and acquire have 650
dierent mappings from the agent patient frame to their shared circum- 651
stance frame (possession transfer). This works as follows. 652
36
We dene the relation between a circumstance and argument frame via 653
a perspectivalizing function. Here are the axioms for what we will call the 654
acquisition function, on which the recipient is agent: 655
(20) (a) acquisition : possession transfer agent patient
(b) agent acquisition = recipient
(c) patient acquisition = possession
(d) source acquisition = donor
656
The rst line denes acquisition as a mapping from the sort possession 657
transfer to the sort agent patient, that is as a mapping from possession 658
transfer eventualities to agent patient eventualities. The mapping is 659
total; that is, each possession transfer is guaranteed to have an agent 660
patient eventuality associated with it. In the second line, the symbol 661
stands for function composition; the composition of the agent function with 662
the acquisition function (written agent acquisition) is the same function 663
(extensionally) as the recipient relation. Thus the ller of the recipient role 664
in a possession transfer must be the same as the ller of the agent role in the 665
associated agent patient eventuality. And so on, for the other axioms. 666
Summing up AVM style: 667
37
(21)
_

_
possession transfer
donor 1
recipient 2
possession 3
_

acquisition
_

_
agent
patient
agent 2
source 1
patient 3
_

_
668
I will call the mapping that makes the donor agent donation. 669
(22)
_

_
possession transfer
donor 1
recipient 2
possession 3
_

donation
_

_
agent
patient
agent 1
goal 2
patient 3
_

_
670
With the acquisition and donation mappings dened, the predicates give 671
and acquire can be dened as compositions with donation and acquisition: 672
give = possession transfer donation

acquire = possession transfer acquisition

673
donation

is an inverse of donation, a function from agent patient even- 674

tualities to possession transfers dened only for those agent patient 675
events related to possession transfers. Composing this with the pos- 676
session transfer predicate makes give a predicate true of those agent 677
38
patient events related to possession transfers, whose agents are donors and 678
whose patients are possessions. The treatment of acquire is parallel but 679
uses the acquisition mappings. For more extensive discussion, see Gawron 680
(2008). 681
Summarizing: 682
a. an argument frame agent patient, with direct consequences for syn- 683
tactic valence (agents become subject, patients direct object, and so on). 684
b. a circumstance frame possession transfer, which captures the circum- 685
stances of possession transfer. 686
c. perspectivalizing functions acquisition and donation which map partici- 687
pants in the circumstances to argument structure. 688
This is the basic picture of perspectivalization. The picture becomes more 689
interesting with a richer example. 690
In the discussion that follows, I assume a commercial transaction frame 691
with at least the following frame elements: 692
39
(23) _

_
commercial
transaction
buyer animate
seller animate
money fungible
goods entity
_

_
693
This is a declaration that various functions from event sorts to truth values 694
and entity sorts exist, a rather austere model for the sort of rich back- 695
grounding function we have assumed for frames. We will see how this model 696
is enriched below. 697
Our picture of proling and perspectivalization can be extended to the 698
more complex cases of commercial transaction predicates with one more 699
composition. For example, we may dene buy as follows: 700
(24) buy = commercial transaction (acquisition goods transfer)

701
What this says is that the relation buy is built in a series of steps, out of 3 702
functions: 703
1. acquisition: the function from possession transfer events to agent patient 704
events already introduced. 705
40
2. goods transfer: a new function from commercial events to possession 706
transfers in which the goods is transferred: 707
_

_
commercial
transaction
buyer 1
seller 2
money 3
goods 4
_

goods-
transfer

_
possession
transfer
recipient 1
donor 2
possession 4
_

_
708
3. The inverse of the composition of goods transfer with acquisition 709
(acquisition goods transfer)

710
is a function from agent patient events to commercial transactions. 711
4. commercial transaction: a sortal predicate true of commercial 712
transactions. 713
5. The predicate buy is therefore true of agent patient events that 714
are related in certain xed ways to a possession transfer and a 715
commercial transaction. 716
The novelty in the denition above is the goods transfer function. We will 717
call this the proling function because it selects the parts of the commercial 718
transaction event which the verb highlights. We will call acquisition 719
41
the function which determines subject and object the perspectivalizing 720
function. The role of the the perspectivalizing function is to select a syntactic 721
realization. 722
A proling function like goods transfer: has two independent motiva- 723
tions: 724
a. It enriches our rather impoverished model of commercial transac- 725
tion. We started out in (23) with little more than the assumption that 726
there were 4 sorted participants we were calling buyer, seller, money, and 727
goods. Now with the assumption of the goods transfer function, a pos- 728
session transfer p is entailed (because the function is total) in which the 729
possession is the goods. Thus goods transfer can be viewed as part of an 730
enriched denition of the commercial transaction frame. There will 731
be other total functions enriching the denition further, for example, a 732
money transfer function of use in dening verbs like pay and collect, in 733
which the money is transferred. 734
b. Both money transfer and goods transfer are projections from com- 735
mercial events to possession transfers; and possession transfer is a frame 736
for which we have a pre-dened perspectivalization, independently moti- 737
vated for other verbs like acquire and get. By composing a commercial 738
event subscene projection with a possession transfer argument projection 739
42
we derive an argument projection for commercial transactions. 740
Thus the good transfer function simultaneously serves knowledge represen- 741
tation needs (a) and valence theory needs (b). 742
There is an analogy between how proling and perspectivalization work 743
and the way the original Minskyan frames work. A Minskyan frame enables 744
the integration of scene components in view with underlying objects by spec- 745
ifying, for example, how the faces of the cube in view relate to the cube as 746
a whole. A Fillmorian perspective enables the integration of the realized 747
elements of a text with an underlying text interpretation by specifying how 748
syntactically realized frame components relate to frames as a whole. In both 749
cases there are operations that mediate between rich representations and a 750
constrained (perspectivalized) representation that belongs to an external 751
representational system. Minskyan rotation operations mediate between 3D 752
representations and the 2D representations of a scene, ultimately necessary 753
because the human retina is a screen. Fillmorian prolings and perspecti- 754
valizations mediate between unlinearized representations in which there is 755
no xed individuation of participants and linearizable argument structure, 756
ultimately necessary because the syntax of human language forces us to 757
linearize participants. 758
Now consider a proling which leaves things out. This is the case of 759
43
spend. 760
(25)
_

_
commercial
transaction
money 1
buyer 2
goods 3
seller 4
_

consumption

_
resource
consumption
resource 1
consumer 2
resource-requirer 3
_

_
761
As discussed in section 2, the verb spend views a commercial transaction as 762
a resource consumption, where resource consumption is the frame 763
used by verbs like waste, lose, use (up), and blow. The proling of the verb 764
spend includes the seller and goods but leaves the seller out. The proling of 765
the verb sell includes the buyer and the goods, as well as the seller. The two 766
subscenes overlap in participants but choose distinct, incompatible event 767
types, which lead to distinct realization possibilities in the syntactic frame. 768
The frame-based picture of commercial transactions is schematized in 769
gure 29.3. 770
The picture on the left shows what we might call the commercial trans- 771
action neighborhood as discussed here. The picture on the right shows that 772
portion of the neighborhood that is activated by buy; the functions used in 773
44
commercial
transaction
possession
transfer
goods
transfer
money
transfer
resource
consumption
consumption
agent_patient
acquisition donation
Buy
commercial
transaction
possession
transfer
goods
transfer
money
transfer
resource
consumption
consumption
agent_patient
acqui-
sition
donation
Fig. 29.3: Left: Lexical network for commercial transaction.
Right: Same network with the perspectivalization chosen by buy in the boxed
area.
45
its denitions are linked by solid lines; the functions left out are in dashes; 774
the boxed regions contains those frames that are used in the denition. If 775
as is suggested in article 108 (Kelter & Kaup) Conceptual knowledge, cate- 776
gorization, and meaning, concepts and word meanings need to be dierent 777
knowledge structures, the picture in gure 29.3 may provide one way of 778
thinking about how they might be related, with the frame nodes playing 779
the role of concepts and a conguration of links between them the role of a 780
word meaning. 781
We have called goods transfer and consumption proling functions. We 782
might equally well have called them subscene roles, because they are func- 783
tions from events to entities. Note that subscene roles dont attribute a 784
xed hierarchical structure to a frame the way do ... cause become .. in 785
Dowtys system attributes a xed structure to causatives of inchoatives. As 786
these examples show, a frame may have subscene roles which carve up its 787
constituents in incompatible ways. Now this may seem peculiar. Shouldnt 788
the roles of a frame dene a xed relation between disjoint entities? I sub- 789
mit that the answer is no. The roles associated with each sort of event 790
are regularities that help us classify an event as of that sort. But such 791
functions are not guaranteed to carve up each event into non-overlapping, 792
hierarchically structured parts. Sometimes distinct roles may select over- 793
46
lapping constituents of events, particularly when independent individuation 794
criteria are not decisive, as when the constituents are collectives, or shape- 795
less globs of stu, or abstract things such as events or event types. Thus we 796
get the cases discussed above like collide,mix, and risk, where dierent ways 797
of proling the frames give us distinct, incompatible sets of roles. We may 798
choose to view the colliders as a single collective entity (X and Y collided), 799
or as two (X collided with Y). We may choose to separate a gure from a 800
ground in the mixing event (14f), or lump them together (mix X and Y), or 801
just view the mixed substance as one (14f). Finally, risks involve an action 802
(13c) and a potential bad consequence (13a), and for a restricted set of cases 803
in which that bad consequence is a loss, a lost thing (13b). 804
What of relations? Formally, in this frame-based picture, we have re- 805
placed relations with event predicates, each of which is dened through some 806
composed set of mappings to a set of events that will be dened only for 807
some xed set of roles. Clearly, for every lexical predicate, there is a corre- 808
sponding relation, namely one dened for exactly the same set of roles as the 809
predicate. Thus in the end the description of the kind of lexical semantic en- 810
tity which interfaces with the combinatorial semantics is not very dierent. 811
However the problems has, I believe, been redened in an interesting way. 812
Traditionally, discussion of the lexical-semantic/syntax interface starts with 813
47
a relation with a predened set of roles. This is the picture for example, 814
that motivates the formulation of Chomskys (1981) -Criterion. However, 815
a major point of frame semantics is that, for many purposes, it is useful to 816
look at a set of relations structured in a particular way. This is the domain 817
of frames. 818
5. Lexicography 819
A word about the application of frames to lexicography is in order. Any set 820
of frames imposes a certain classicational scheme on the lexicon. Other ex- 821
amples of such a classicational scheme are Rogets Thesaurus, Longmans 822
valence classes, and Wordnet (Fellbaum 1998). Frames dier from all three 823
in that they are not primarily oriented either to the task of synonym-classes 824
or syntactic frame classes. One expects to nd synonyms and antonyms in 825
the same frame, of course, and many examples of valence similarity, but 826
neither trend will be a rule. As we saw in section 2, near synonyms like land 827
and ground may belong to dierent frames, and understanding those frames 828
is critical to proper usage. As we saw in our investigations of proling and 829
perspective, dierences of both kinds may result in very dierent valence 830
options for verbs from the same frame. The value of the frame idea for lex- 831
icography is that it seems the most promising idea if the goal is to organize 832
48
words according to usage. This of course is a hypothesis. FrameNet (Fill- 833
more & Baker 2000) is a test of that hypothesis. Accordingly, frame entries 834
are connected with rich sets of examples gleaned from the British National 835
Corpus illustrating frame element realizations in a variety of syntactic con- 836
texts. Interested readers will nd a tour of the web site far more persuasive 837
than any discussion here. 838
6. Discourse understanding 839
In this section I propose to raise the issue of frames in discourse under- 840
standing, not to try to give the subject an adequate treatment, for which 841
there is no space, but to talk a bit about how the role of frames in discourse 842
understanding is related to their role in interpreting signs. 843
Let us return to the example of verbs conventionally connected with 844
eects caused by movement: 845
(26) a. John broke the glass against the wall. 846
b. # John killed the cockroach against the wall. 847
It is at least arguably the case that this contrast can be made without the 848
help of a lexical stipulation. If movement can be a default or at least a 849
highly prototypical way of breaking something, and not a highly prototypi- 850
49
cal way of killing something, then something like the default logic of Asher 851
& Lascarides (1995) or abduction as in Hobbs et al. (1993), both of which 852
have been applied successfully to a number of problems of discourse inter- 853
pretation, could infer causality in (a) and not in (b). However, this still 854
falls somewhat short of predicting the genuine oddity of (b). Notice, too, 855
that when discourse coherence alone is at issue, both causality inferences go 856
through: 857
(27) a. The glass was hurled against the wall and broke. 858
b. The cockroach was hurled against the wall and died. 859
Thus the defaults at play in determining matters of valence dier from 860
those in discourse. We can at least describe the contrasts in (26) not 861
explain it by saying movement is an optional component of the breaking 862
frame through which the denotation of the verb break is dened, and not a 863
component of the killing frame; or in terms of the formal picture of section 864
4: Within the conventional lexical network linking frames in English there 865
is a partial function from breaking events to movement subscenes; there is 866
no such function for killing events. 867
In contrast Fillmores (1985: 232) discussed in section 2.1: 868
(28) We never open our presents until morning. 869
50
The point of this example was that it evoked Christmas without containing 870
a single word specic to Christmas. How might an automatic interpretation 871
system simulate what is going on for human understanders? Presumably by 872
a kind of application of Occams razor. There is one and only one frame 873
that explains both the presence of presents and the custom of waiting until 874
morning, and that is the Christmas frame. Thus the assumption that gets 875
us the most narrative bang for the buck is Christmas. In this case the frame 876
has to be evoked by dynamically assembling pieces of information activated 877
in this piece of discourse. 878
These two examples show that frames will function dierently in a theory 879
of discourse understanding than they will in a theory of sign-meanings in 880
at least two ways. They will require a dierent notion of default, and they 881
will need to resort to dierent inferencing strategies, such as inference to the 882
most economical explanation. 883
7. Conclusion 884
The logical notion of a relation, which preserves certain aspects of the lin- 885
earization syntax forces on us, has at times appeared to oer an attractive 886
account of what we grasp when we grasp sign meanings. But the data we 887
have been looking at in this brief excursion into frame semantics has pointed 888
51
another way. Lexical senses seem to be tied to the same kind schemata that 889
organize our perceptions and interpretations of the social and physical world. 890
In these schemata participants are neither linearized nor uniquely individ- 891
uated, and the mapping into the linearized regime of syntax is constrained 892
but underdetermined. We see words with options in what their exact par- 893
ticipants are and how they are realized. Frames oer a model that is both 894
specic enough and exible enough to accommodate these facts, while of- 895
fering the promise of a rm grounding for lexicographic description and an 896
account of text understanding. 897
8. References 898
Asher, Nicholas & Alex Lascarides 1995. Lexical disambiguation in a dis- 899
course context. Journal of Semantics 12(1), 69108. 900
Baker, Collin & Charles J. Fillmore 2001. Frame Semantics for text under- 901
standing. In: Proceedings of WordNet and Other Lexical Resources 902
Workshop. Pittsburgh, PA: North American American Association of 903
Computational Linguistics, 34. 904
Baker, Collin, Charles J. Fillmore & John B. Lowe 1998. The Berkeley 905
FrameNet project. In: B.T.S. Atkins & A. Zampolli (eds.). Proceed- 906
ings of the Joint Conference of the International Committee on Com- 907
52
putational Linguistics and the Association for Computational Linguis- 908
tics, Vol. 1. Montreal: Association for Computational Linguistics, 909
8690. 910
Barsalou, Lawrence 1992. Frames, concepts, and conceptual elds. In: 911
A. Lehrer & E. Kittay (eds.). Frames, Fields, and Contrasts. New 912
Essays in Semantic and Lexical Organization. Hillsdale, NJ: Lawrence 913
Erlbaum, 2174. 914
Barsalou, Lawrence 1999. Perceptual symbol systems. Behavioral and Brain 915
Sciences 22, 577660. 916
Bartlett, Frederick C. 1932. Remembering. A Study in Experimental and 917
Social Psychology. Cambridge: Cambridge University Press. 918
Berlin, Isaiah 1957. The Hedgehog and the Fox. New York: New American 919
Library. Reprinted in Berlin (1997). 920
Berlin, Isaiah 1997. The Proper Study of Mankind. New York: Farrar, 921
Straus, and Giroux. 922
Boas, Hans C. 2001. Frame Semantics as a framework for describing pol- 923
ysemy and syntactic structures of English and German motion verbs 924
in contrastive computational lexicography. In: P. Rayson, A. Wilson, 925
T. McEnery, A. Hardie & S. Khoja (eds.). Proceedings of the Corpus 926
Linguistics 2001 Conference. Lancaster, UK: University Centre for 927
53
Computer Corpus Research on Language, 6473. 928
Boas, Hans C. 2005. Semantic frames as interlingual representations for 929
multilingual databases. International Journal of Lexicography 18(4), 930
445478. 931
Carpenter, Bob 1992. The Logic of Typed Feature Structures. With Ap- 932
plications to Unication Grammars, Logic Programs and Constraint 933
Resolution. Cambridge, MA: Cambridge University Press. 934
Chang, Nancy, Srini Narayanan & Miriam R.L. Petruck 2002a. From frames 935
to inference. In: Proceedings of the First International Workshop 936
on Scalable Natural Language Understanding. Heidelberg, Germany: 937
International Committee on Computational Linguistics, 478484. 938
Chang, Nancy, Srini Narayanan & Miriam R.L. Petruck 2002b. Putting 939
frames in perspective. In: Proceedings of the Nineteenth International 940
Conference on Computational Linguistics. Taipei, Taiwan: Interna- 941
tional Committee on Computational Linguistics, 231237. 942
Chomsky, Noam 1981. Lectures on Government and Binding. Dordrecht: 943
Foris. 944
Coseriu, Eugenio 1967. Lexikalische Solidaritaten. Poetica 1, 293303. 945
Davidson, Donald 1967. The logical form of action sentences. In: N. Rescher 946
(ed.). The Logic of Decision and Action. University of Pittsburgh 947
54
Press, 8194. Reprinted in Davidson (1980), chapter 6. 948
Davidson, Donald 1980. Essays on Actions and Events. Oxford University 949
Press. 950
Dowty, David R. 1979. Word Meaning and Montague Grammar. Boston: 951
Reidel. 952
Dowty, David R. 1991. Thematic proto roles and argument selection. Lan- 953
guage 67, 547619. 954
Fellbaum, Christiane (ed.) 1998. WordNet. An Electronic Lexical Database. 955
Cambridge, MA: MIT Press. 956
Fillmore, Charles J. 1975. An alternative to checklist theories of mean- 957
ing. In: C. Cogen, H. Thompson, G. Thurgood, K. Whistler & J. 958
Wright (eds.). Proceedings of the First Annual Meeting of the Berke- 959
ley Linguistics Society. University of California at Berkeley: Berkeley 960
Linguistics Society, 123131. 961
Fillmore, Charles J. 1976. Topics in lexical semantics. In: R. Cole (ed.). 962
Current Issues in Linguistic Theory. Indiana University Press, 76 963
138. 964
Fillmore, Charles J. 1977a. The case for case reopened. In: P. Cole & J. 965
Sadock (eds.). Syntax and Semantics, Vol. 8: Grammatical Relations. 966
New York: Academic Press, 5981. 967
55
Fillmore, Charles J. 1977b. Scenes-and-frames Semantics. Linguistic struc- 968
tures processing. In: A. Zampolli (ed.). Linguistic Structures Pro- 969
cessing. Amsterdam: North Holland, 5581. 970
Fillmore, Charles J. 1978. On the organization of semantic information in 971
the lexicon. In: Papers from the Parasession on the Lexicon. Univer- 972
sity of Chicago: Chicago Linguistics Society, 148173. 973
Fillmore, Charles J. 1982. Frame Semantics. In: Linguistic Society of Korea 974
(ed.). Linguistics in the Morning Calm. Seoul: Hanshin, 111137. 975
Fillmore, Charles J. 1985. Frames and the semantics of understanding. 976
Quaderni di Semantica 6(2), 222254. 977
Fillmore, Charles J. & B.T.S Atkins 1994. Starting where the dictionaries 978
stop. The challenge for computational lexicography. In: B.T.S Atkins 979
& A. Zampolli (eds.). Computational Approaches to the Lexicon. Ox- 980
ford: Clarendon Press. 981
Fillmore, Charles J. & B.T.S Atkins 1998. FrameNet and lexicographic rel- 982
evance. In: Proceedings of the First International Conference on Lan- 983
guage Resources and Evaluation. Granada: International Committee 984
on Computational Linguistics, 2830. 985
Fillmore, Charles J. & Collin Baker 2000. FrameNet, 986
http://www.icsi.berkeley.edu/framenet, October 13, 2008. 987
56
Gawron, Jean Mark 1983. Lexical Representations and the Semantics of 988
Complementation. New York: Garland. 989
Gawron, Jean Mark 2008. Circumstances and Perspective. The Logic of 990
Argument Structure. http://repositories.cdlib.org/ucsdling/sdlp3/1, 991
October 13, 2008. 992
Geckeler, Horst 1971. Strukturelle Semantik und Wortfeldtheorie. M unchen: 993
Fink. 994
Hobbs, Jerry R., Mark Stickel, Douglas Appelt & Paul Martin 1993. Inter- 995
pretation as abduction. Articial Intelligence 63, 69142. 996
Jackendo, Ray S. 1990. Semantic Structures. Cambridge, MA: MIT Press. 997
Koka, Kurt 1963. Principles of Gestalt Psychology. New York: Harcourt, 998
Brace, and World. 999
Lako, George 1972. Linguistics and natural logic. In: D. Davidson & G. 1000
Harman (eds.). Semantics of Natural Language. Dordrecht: Reidel, 1001
545665. 1002
Lako, George 1983. Category. An essay in cognitive linguistics. In: Lin- 1003
guistic Society of Korea (ed.). Linguistics in the Morning Calm. Seoul: 1004
Hanshin, 139194. 1005
Lako, George & Mark Johnson 1980. Metaphors We Live By. Chicago: 1006
University of Chicago Press. 1007
57
Langacker, Ronald 1984. Active zones. In: Proceedings of the Tenth Annual 1008
Meeting of the Berkeley Linguistics Society. University of California 1009
at Berkeley: Berkeley Linguistics Society, 172188. 1010
Lehrer, Adrienne & Eva Kittay (eds.) 1992. Frames, Fields, and Con- 1011
trasts. New Essays in Semantic and Lexical Organization. Hillsdale, 1012
NJ: Lawrence Erlbaum. 1013
Levin, Beth 1993. English Verb Classes and Alternations. Chicago: Univer- 1014
sity of Chicago Press. 1015
Mill, John Stuart 1847. A System of Logic. New York: Harper and Brothers. 1016
Minsky, Marvin 1975. A framework for representing knowledge. In: P. 1017
Winston (ed.). The Psychology of Computer Vision. McGraw-Hill, 1018
211277. 1019
http://web.media.mit.edu/minsky/papers/Frames/frames.html, Oc- 1020
tober 13, 2008. 1021
Parsons, Terence 1990. Events in the Semantics of English. Cambridge, 1022
MA: MIT Press. 1023
Pustejovsky, James 1995. The Generative Lexicon. Cambridge, MA: MIT 1024
Press. 1025
Rounds, William C. 1997. Feature logics. In: J. van Benthem & A. ter 1026
Meulen (eds.). Handbook of Logic and Language. Cambridge, MA: 1027
58
Elsevier, 475533. 1028
Rumelhart, Donald 1980. Schemata. The building blocks of cognition. In: 1029
R.J. Spiro, B.C. Bruce & W.F. Brewer (eds.). Theoretical Issues in 1030
Reading Comprehension. Hillsdale, NJ: Lawrence Erlbaum, 3358. 1031
Schank, Roger C. & R.P. Abelson 1977. Scripts, Plans, Goals, and Under- 1032
standing. Hillsdale, NJ: Lawrence Erlbaum. 1033
Smolka, Gert 1992. Feature constraint logics for unication grammars. 1034
Journal of Logic Programming 12, 5187. 1035
Trier, Jost 1971. Aufs atze und Votr age zur Wortfeldtheorie. The Hague: 1036
Mouton. 1037
Weisgerber, J. Leo 1962. Grundz uge der inhaltsbezogenen Grammatik. D usseldorf: 1038
Schwann. 1039
Jean-Mark Gawron, San Diego (USA) 1040
59